Merge pull request #4924 from jtrmal/python_and_apple

Fix reported issues w.r.t python2.7 and some apple silicone quirks

Author: danpovey

src/Makefile

@@ -62,14 +62,16 @@ endif
 
 # Don't call rm -rf.
 rmlibdir:
+ifeq ($(KALDI_FLAVOR), dynamic)
 ifneq ($(KALDILIBDIR), )
-	-rm -f $(KALDILIBDIR)/*{.so,.a,.o}
+	-rm -f $(KALDILIBDIR)/*{.so,.a,.o,.dylib}
 	-rmdir 2>/dev/null $(KALDILIBDIR); true
 else
 # KALDILIBDIR might have been unset because of reconfigure. Do a best guess.
 	@echo "Something seems wrong. Please re-run configure."
 	@echo "I will continue but the cleanup might not be complete."
 endif
+endif
 
 kaldi.mk:
 	@echo "ERROR: kaldi.mk does not exist; run ./configure first.";

src/configure

@@ -1150,6 +1150,8 @@ elif [ "`uname`" == "Darwin" ]; then
       cat makefiles/darwin_clapack.mk >> kaldi.mk
       echo "Warning (CLAPACK): this part of the configure process is not properly tested and may not work."
       echo "Successfully configured for Darwin with CLAPACK libs from $CLAPACKROOT"
+    elif [ "$(uname -m)" == "arm64" ]; then
+      cat makefiles/darwin_arm64.mk >> kaldi.mk
     else
       cat makefiles/darwin.mk >> kaldi.mk
     fi

src/cudamatrix/cu-array.h

@@ -105,13 +105,12 @@ class CuArrayBase {
  protected:
   /// Default constructor: make it protected so the user cannot
   /// instantiate this class.
-  CuArrayBase<T>(): data_(NULL), dim_(0) { }
+  CuArrayBase(): data_(NULL), dim_(0) { }
 
 
   T *data_;  ///< GPU data pointer (if GPU not available,
              ///< will point to CPU memory).
   MatrixIndexT dim_;     ///< dimension of the vector
-
 };
 
 /**
@@ -123,22 +122,21 @@ class CuArrayBase {
 template<typename T>
 class CuArray: public CuArrayBase<T> {
  public:
-
   /// Default constructor, initialized data_ to NULL and dim_ to 0 via
   /// constructor of CuArrayBase.
-  CuArray<T>() { }
+  CuArray() { }
 
   /// Constructor with memory initialisation.  resize_type may be kSetZero or
   /// kUndefined.
-  explicit CuArray<T>(MatrixIndexT dim, MatrixResizeType resize_type = kSetZero)
+  explicit CuArray(MatrixIndexT dim, MatrixResizeType resize_type = kSetZero)
      { Resize(dim, resize_type); }
 
   /// Constructor from CPU-based int vector
-  explicit CuArray<T>(const std::vector<T> &src) { CopyFromVec(src); }
+  explicit CuArray(const std::vector<T> &src) { CopyFromVec(src); }
 
   /// Copy constructor.  We don't make this explicit because we want to be able
   /// to create a std::vector<CuArray>.
-  CuArray<T>(const CuArray<T> &src) { CopyFromArray(src); }
+  CuArray(const CuArray<T> &src) { CopyFromArray(src); }
 
   /// Destructor
   ~CuArray() { Destroy(); }
@@ -172,7 +170,6 @@ class CuArray: public CuArrayBase<T> {
   /// I/O
   void Read(std::istream &is, bool binary);
   void Write(std::ostream &is, bool binary) const;
-
 };
 
 
@@ -182,7 +179,7 @@ class CuSubArray: public CuArrayBase<T> {
   /// Constructor as a range of an existing CuArray or CuSubArray.  Note: like
   /// similar constructors in class CuVector and others, it can be used to evade
   /// 'const' constraints; don't do that.
-  explicit CuSubArray<T>(const CuArrayBase<T> &src,
+  explicit CuSubArray(const CuArrayBase<T> &src,
                          MatrixIndexT offset, MatrixIndexT dim);
 
   /// Construct from raw pointers

src/cudamatrix/cu-matrix.h

@@ -250,7 +250,7 @@ class CuMatrixBase {
   template<typename OtherReal>
   void CopyFromTp(const CuTpMatrix<OtherReal> &M,
                   MatrixTransposeType trans = kNoTrans);
-  
+
   // This function will copy from source rows (start_range, end_range]
   // if the range is outside of the clamped region then the clamped
   // row will be replicated across the out of range areas
@@ -307,9 +307,9 @@ class CuMatrixBase {
   void PowAbs(const CuMatrixBase<Real> &src, Real power, bool include_sign=false);
 
   void Floor(const CuMatrixBase<Real> &src, Real floor_val);
-  
+
   void Ceiling(const CuMatrixBase<Real> &src, Real ceiling_val);
-  
+
   /// This is equivalent to running:
   /// Floor(src, lower_limit);
   /// Ceiling(src, upper_limit);
@@ -320,7 +320,7 @@ class CuMatrixBase {
   /// (x < 0 ? exp(x) : x + 1).  This function is used
   /// in our RNNLM training.
   void ExpSpecial(const CuMatrixBase<Real> &src);
-  
+
   /// Softmax nonlinearity
   /// Y = Softmax(X) : Yij = e^Xij / sum_k(e^Xik), done to each row,
   /// with attention to avoiding  overflow or underflow.
@@ -333,7 +333,7 @@ class CuMatrixBase {
   /// Supports in-place operation (i.e. this == &src).
   void LogSoftMaxPerRow(const CuMatrixBase<Real> &src);
 
-  
+
   /// Apply the function y = log(1 + exp(x)), to each element.
   /// Note: the derivative of this function is the sigmoid function.
   /// This is like a soft ReLU.
@@ -439,23 +439,23 @@ class CuMatrixBase {
     this -> Pow(*this, power);
   };
 
-  
+
   inline void ApplyPowAbs(Real power, bool include_sign=false) {
     this -> PowAbs(*this, power, include_sign);
   };
-  
+
   inline void ApplyHeaviside() {
     this -> Heaviside(*this);
   };
-  
+
   inline void ApplyFloor(Real floor_val) {
     this -> Floor(*this, floor_val);
   };
-  
+
   inline void ApplyCeiling(Real ceiling_val) {
     this -> Ceiling(*this, ceiling_val);
   };
-  
+
   inline void ApplyExp() {
     this -> Exp(*this);
   };
@@ -924,7 +924,7 @@ class CuSubMatrix: public CuMatrixBase<Real> {
 
   /// This type of constructor is needed for Range() to work [in CuMatrix base
   /// class]. Cannot make it explicit or that breaks.
-  inline CuSubMatrix<Real> (const CuSubMatrix &other):
+  inline CuSubMatrix(const CuSubMatrix &other):
   CuMatrixBase<Real> (other.data_, other.num_rows_, other.num_cols_,
                       other.stride_) {}
  private:

src/cudamatrix/cu-tp-matrix.h

@@ -48,18 +48,18 @@ class CuTpMatrix : public CuPackedMatrix<Real> {
   CuTpMatrix() : CuPackedMatrix<Real>() {}
   explicit CuTpMatrix(MatrixIndexT r, MatrixResizeType resize_type = kSetZero)
       : CuPackedMatrix<Real>(r, resize_type) {}
-  
-  explicit CuTpMatrix<Real>(const TpMatrix<Real> &orig)
+
+  explicit CuTpMatrix(const TpMatrix<Real> &orig)
       : CuPackedMatrix<Real>(orig) {}
   // This constructor lacks the "explicit" keyword so that
   // we can include this class in std::vector.
-  CuTpMatrix<Real>(const CuTpMatrix<Real> &orig)
+  CuTpMatrix(const CuTpMatrix<Real> &orig)
       : CuPackedMatrix<Real>(orig) {}
-  
-  explicit CuTpMatrix<Real>(const CuMatrixBase<Real> &orig,
+
+  explicit CuTpMatrix(const CuMatrixBase<Real> &orig,
                             MatrixTransposeType trans = kNoTrans);
 
-  
+
   ~CuTpMatrix() {}
 
   void CopyFromMat(const CuMatrixBase<Real> &M,
@@ -70,12 +70,12 @@ class CuTpMatrix : public CuPackedMatrix<Real> {
   }
   void CopyFromTp(const TpMatrix<Real> &other) {
     CuPackedMatrix<Real>::CopyFromPacked(other);
-  }  
+  }
   void Cholesky(const CuSpMatrix<Real>& Orig);
   void Invert();
 
   CuTpMatrix<Real> &operator = (const CuTpMatrix<Real> &in);
-  
+
  protected:
   inline const TpMatrix<Real> &Mat() const {
     return *(reinterpret_cast<const TpMatrix<Real>* >(this));

src/cudamatrix/cu-vector.h

@@ -243,7 +243,7 @@ class CuVectorBase {
 
   /// Default constructor: make it protected so the user cannot
   /// instantiate this class.
-  CuVectorBase<Real>(): data_(NULL), dim_(0) { }
+  CuVectorBase(): data_(NULL), dim_(0) { }
 
   Real *data_; ///< GPU data pointer (or regular data pointer
                ///< if CUDA is not compiled in or we have no GPU).

src/makefiles/darwin_arm64.mk

@@ -0,0 +1,36 @@
+# Darwin (macOS) configuration
+
+ifndef DOUBLE_PRECISION
+$(error DOUBLE_PRECISION not defined.)
+endif
+ifndef OPENFSTINC
+$(error OPENFSTINC not defined.)
+endif
+ifndef OPENFSTLIBS
+$(error OPENFSTLIBS not defined.)
+endif
+
+CXXFLAGS = -std=c++14 -I.. -I$(OPENFSTINC) -O1 $(EXTRA_CXXFLAGS) \
+           -Wall -Wno-sign-compare -Wno-unused-local-typedefs \
+           -Wno-deprecated-declarations -Winit-self \
+           -DKALDI_DOUBLEPRECISION=$(DOUBLE_PRECISION) \
+           -DHAVE_EXECINFO_H=1 -DHAVE_CXXABI_H -DHAVE_CLAPACK \
+           -pthread \
+           -g # -O0 -DKALDI_PARANOID
+
+ifeq ($(KALDI_FLAVOR), dynamic)
+CXXFLAGS += -fPIC
+endif
+
+# Compiler specific flags
+COMPILER = $(shell $(CXX) -v 2>&1)
+ifeq ($(findstring clang,$(COMPILER)),clang)
+# Suppress annoying clang warnings that are perfectly valid per spec.
+CXXFLAGS += -Wno-mismatched-tags
+else ifeq ($(findstring GCC,$(COMPILER)),GCC)
+# Allow implicit conversions between vectors.
+CXXFLAGS += -flax-vector-conversions
+endif
+
+LDFLAGS = $(EXTRA_LDFLAGS) $(OPENFSTLDFLAGS) -g
+LDLIBS = $(EXTRA_LDLIBS) $(OPENFSTLIBS) -framework Accelerate -lm -lpthread -ldl

src/matrix/kaldi-matrix.h

@@ -1006,11 +1006,11 @@ class SubMatrix : public MatrixBase<Real> {
             MatrixIndexT num_cols,
             MatrixIndexT stride);
 
-  ~SubMatrix<Real>() {}
+  ~SubMatrix() {}
 
   /// This type of constructor is needed for Range() to work [in Matrix base
   /// class]. Cannot make it explicit.
-  SubMatrix<Real> (const SubMatrix &other):
+  SubMatrix(const SubMatrix &other):
   MatrixBase<Real> (other.data_, other.num_cols_, other.num_rows_,
                     other.stride_) {}
 

src/matrix/qr.cc

@@ -57,7 +57,7 @@ void House(MatrixIndexT dim, const Real *x, Real *v, Real *beta) {
     if (max_x == 0.0) max_x = 1.0;
     s = 1.0 / max_x;
   }
-  
+
   Real sigma = 0.0;
   v[0] = 1.0;
   for (MatrixIndexT i = 1; i < dim; i++) {
@@ -73,7 +73,7 @@ void House(MatrixIndexT dim, const Real *x, Real *v, Real *beta) {
       v[0] = x1 - mu;
     } else {
       v[0] = -sigma / (x1 + mu);
-      KALDI_ASSERT(KALDI_ISFINITE(v[dim-1]));      
+      KALDI_ASSERT(KALDI_ISFINITE(v[dim-1]));
     }
     Real v1 = v[0];
     Real v1sq = v1 * v1;
@@ -155,11 +155,11 @@ void HouseBackward(MatrixIndexT dim, const Real *x, Real *v, Real *beta) {
    with packed lower-triangular matrices to do it this way.  There's also
    a shift from one-based to zero-based indexing, so the index
    k is transformed k -> n - k, and a corresponding transpose...
-   
+
    Let the original *this be A.  This algorithms replaces *this with
    a tridiagonal matrix T such that T = Q A Q^T for an orthogonal Q.
    Caution: Q is transposed vs. Golub and Van Loan.
-   If Q != NULL it outputs Q. 
+   If Q != NULL it outputs Q.
 */
 template<typename Real>
 void SpMatrix<Real>::Tridiagonalize(MatrixBase<Real> *Q) {
@@ -195,7 +195,7 @@ void SpMatrix<Real>::Tridiagonalize(MatrixBase<Real> *Q) {
     if (Q != NULL) { // C.f. Golub, Q is H_1 .. H_n-2... in this
       // case we apply them in the opposite order so it's H_n-1 .. H_1,
       // but also Q is transposed so we really have Q = H_1 .. H_n-1.
-      // It's a double negative.    
+      // It's a double negative.
       // Anyway, we left-multiply Q by each one.  The H_n would each be
       // diag(I + beta v v', I) but we don't ever touch the last dims.
       // We do (in Matlab notation):
@@ -309,7 +309,7 @@ void QrStep(MatrixIndexT n,
     if (k < n-2) {
       // Next is the elements (k+2, k) and (k+2, k-1), to be rotated, again
       // backwards.
-      Real &elem_kp2_k = z, 
+      Real &elem_kp2_k = z,
           &elem_kp2_kp1 = off_diag[k+1];
       // Note: elem_kp2_k == z would start off as zero because it's
        // two off the diagonal, and not been touched yet.  Therefore
@@ -338,7 +338,7 @@ void QrInternal(MatrixIndexT n,
   MatrixIndexT counter = 0, max_iters = 500 + 4*n, // Should never take this many iters.
       large_iters = 100 + 2*n;
   Real epsilon = (pow(2.0, sizeof(Real) == 4 ? -23.0 : -52.0));
-  
+
   for (; counter < max_iters; counter++) { // this takes the place of "until
                                            // q=n"... we'll break out of the
                                            // loop when we converge.
@@ -356,7 +356,7 @@ void QrInternal(MatrixIndexT n,
         off_diag[i] = 0.0;
     }
     // The next code works out p, q, and npq which is n - p - q.
-    // For the definitions of q and p, see Golub and Van Loan; we 
+    // For the definitions of q and p, see Golub and Van Loan; we
     // partition the n dims into pieces of size (p, n-p-q, q) where
     // the part of size q is diagonal and the part of size n-p-p is
     // "unreduced", i.e. has no zero off-diagonal elements.
@@ -392,7 +392,7 @@ void QrInternal(MatrixIndexT n,
     } else {
       QrStep(npq, diag + p, off_diag + p,
              static_cast<MatrixBase<Real>*>(NULL));
-    }      
+    }
   }
   if (counter == max_iters) {
     KALDI_WARN << "Failure to converge in QR algorithm. "
@@ -490,7 +490,7 @@ void SpMatrix<Real>::TopEigs(VectorBase<Real> *s, MatrixBase<Real> *P,
     r.AddSpVec(1.0, S, Q.Row(d), 0.0);
     // r = S * q_d
     MatrixIndexT counter = 0;
-    Real end_prod;
+    Real end_prod = 0;
     while (1) { // Normally we'll do this loop only once:
       // we repeat to handle cases where r gets very much smaller
       // and we want to orthogonalize again.
@@ -528,11 +528,11 @@ void SpMatrix<Real>::TopEigs(VectorBase<Real> *s, MatrixBase<Real> *P,
     }
   }
 
-  Matrix<Real> R(lanczos_dim, lanczos_dim);  
+  Matrix<Real> R(lanczos_dim, lanczos_dim);
   R.SetUnit();
   T.Qr(&R); // Diagonalizes T.
   Vector<Real> s_tmp(lanczos_dim);
-  s_tmp.CopyDiagFromSp(T);  
+  s_tmp.CopyDiagFromSp(T);
 
   // Now T = R * diag(s_tmp) * R^T.
   // The next call sorts the elements of s from greatest to least absolute value,
@@ -544,7 +544,7 @@ void SpMatrix<Real>::TopEigs(VectorBase<Real> *s, MatrixBase<Real> *P,
   SubMatrix<Real> Rsub(R, 0, eig_dim, 0, lanczos_dim);
   SubVector<Real> s_sub(s_tmp, 0, eig_dim);
   s->CopyFromVec(s_sub);
-      
+
   // For working out what to do now, just assume the other eigenvalues were
   // zero.  This is just for purposes of knowing how to get the result, and
   // not getting things wrongly transposed.